Threat Intelligence

When an AI Agent Acts as You: Lessons From a Tool That Broke Every Security Rule

The Hacker News · 11 Nov 2026
Key Takeaway Before letting any AI agent work inside a staff member's logged-in session, check whether your tools can tell the agent apart from the person, and limit what it can reach.

Token, a security company, says it normally advises customers that AI agents need their own identities, access that fits the job, and traceability separate from any human's. Then it built an agent called Rehearsal that broke all three by default. Rehearsal answers questions inside the Token platform by doing the work: it opens pages in the browser tab, sets filters and fills in forms.

The problem is that Rehearsal runs inside the user's authenticated session and inherits their reach. To the server behind the page, the agent's clicks look the same as the user's, because nothing in a session cookie shows who is actually driving. The application may see the user, but not the agent, the delegation, or the intended scope of the task.

Because Token owns both the agent and the app, it could close these gaps from the inside. For example, Rehearsal tells the platform that it is the one acting. But the company notes this only works because it controls both sides. A browser agent driving a CRM or cloud console has no such arrangement: the app cannot tell the agent is there, and the agent cannot tell which field holds a secret or which button deletes a record. Token argues the way forward is to give agents their own lanes, separate from human users. The web has started building these, though not yet in a holistic way. Standards are coming, but teams can take steps today.

Running AI agents? Eight controls for securing them, mapped to the ISM ->

Summarised by CISO AI from The Hacker News, written with Claude Sonnet 5.5. We link back to every original so you can read it yourself.